WITHDRAWN: The Consciousness Bottleneck: Systematic Scaling Reveals a Sweet Spot for Mirror Neurons in Recurrent Neural Networks

Mirror neurons—units that fire both when an agent performs an action and when it observes the same action performed by others—have been implicated in social cognition, imitation, and self–other understanding (Gallese et al., 2004). But it remains unclear whether they are a biologically specialized mechanism or a generic computational consequence of learning. Here we present a minimal, fully controlled computational study showing that mirror-like units emerge robustly from a simple architectural requirement: a shared latent bottleneck trained bidirectionally on forward and inverse prediction (Wolpert, Ghahramani, & Jordan, 1995; Kawato, 1999). We train recurrent neural networks (GRUs) in a synthetic sensorimotor environment with two self-supervised objectives: predicting sensory consequences from actions (forward model) and inferring actions from sensory streams (inverse model). When both tasks share a single recurrent core, we observe substantial unitwise alignment between perform and observe modes. Across 2,000 sequences and 20 epochs in a canonical setting (hidden size 96), 25–30% of units exhibit |r|>0.75 between perform and observe activations, ~ 35–40% have |r|>0.6, and peak units reach r ≈ 0.9. Ablations with separate cores, forward-only objectives, or shuffled observation labels abolish the effect, establishing its architectural origin. We then conduct, to our knowledge, the first systematic capacity-scaling study of mirror-like representations. Varying GRU hidden size from 4 to 512 and evaluating with strict controls (independent test sequences, untrained baselines), we find a pronounced “consciousness bottleneck” sweet spot at intermediate capacity (h ≈ 32–48). At this capacity range, four metrics co-optimize: (i) true mirror effect Δr (difference between shared- and separate-core maximum |r|), (ii) information-bottleneck compression ratio between perform and observe states, (iii) self/other decoder accuracy, and (iv) critical-period-like developmental dynamics. Beyond the sweet spot, larger models achieve lower prediction error but weaker mirror alignment and reduced compression, breaking the usual monotonic ML scaling laws (Kaplan et al., 2020; Wei et al., 2022). These results support a view of mirror neurons not as biological quirks, but as computational necessities induced by shared forward–inverse architectures under bottleneck pressure. More broadly, they suggest that consciousness-adjacent properties—self–other mapping, compression, and critical periods—may arise from representational constraints rather than sheer scale (Friston, 2010; Dehaene, Kerszberg, & Changeux, 1998; Bengio, 2017), with implications for neuroscience, developmental psychology, robotics, and a nascent science of “robopsychology.”